Research Article

---
name: research-article
description: Research a topic to map the state of the art, gathering and triaging sources before writing an article. Produces a research brief with organized sources, key findings, and identified gaps that feeds into create-article.
---

# Research Article

Given a topic, discovers what the state of the art is by searching broadly, triaging sources, deep-reading the strongest ones, and synthesizing the field.
The output is a research brief that becomes the input to `/create-article`, so the writing step can focus on craft rather than discovery.

## Prerequisites

- **Topic**: The subject to investigate (e.g., "vector databases", "test-time compute in LLMs", "CRDTs for collaborative editing")
- **Scope** (optional but recommended): The specific angle or question, time bounds (e.g., "last 2 years"), and what "state of the art" means here (best accuracy, fastest, most adopted, most recent breakthrough)
- **Intended audience** (optional): Helps calibrate which sources and framing are relevant
- **Seed sources** (optional): Any URLs, papers, or files the user already knows matter

## Steps

1. Clarify the topic and scope.
   State the specific research question in one sentence.
   Define what "state of the art" means for this topic (top benchmark score, dominant method, industry standard, latest breakthrough).
   Note any time bounds or constraints.
2. Search broadly.
   Search multiple source types (see Search Strategy below).
   Start with surveys, review papers, and benchmark leaderboards to map the field quickly before going deep.
3. Build a candidate source list.
   For each result, capture title, URL, type, year, and a one-line relevance note.
   Prioritize primary sources (papers, official docs, benchmarks) over secondary commentary, but keep influential blog posts and talks that shape practice.
4. Triage.
   Rank candidates by relevance, credibility, and recency.
   Score credibility from concrete authority signals, not impressions:
   - Papers: a peer-reviewed venue and citations beat an uncited preprint; note the sample size.
   - Reports and standards: the publishing organization and its methodology (for example NBER, DORA, OpenTelemetry).
   - Repositories: stars, forks, and recent activity, and flag a small or obscure repo even when it states a useful number.
   - Vendor content: a company with a product in the category is self-interested; keep it, but label it and never present it as independent.
   - Blogs: a named author and no sales pitch; a services-company SEO post with no author is Weak.
   Record an authority signal and a credibility rating (Strong, Moderate, or Weak, with a qualifier such as "Strong (small n)") for every source you keep.
   Keep the strongest 8-15 to deep-read.
   Drop outdated or low-quality sources, but record why so the exclusion is deliberate.
5. Deep-read the top sources.
   Fetch and read each one.
   Extract: the problem it addresses, its approach, its key results and numbers, and how it compares to prior work.
   Verify claims against the source rather than paraphrasing from memory or a summary.
6. Map the field.
   Group the sources into the main approaches.
   Identify what currently defines the state of the art (top scores, dominant technique, widely adopted standard).
   Note where the field agrees and where it actively disagrees.
7. Surface gaps and open questions.
   What is unsolved, contested, or just emerging?
   These tensions are what make an article worth reading.
8. Synthesize.
   Write the state-of-the-art summary.
   Recommend an angle for the article: the single core insight or argument the research best supports.
9. Produce the research brief using the output format below.

## Search Strategy

Cover several source types so the result is not skewed by one medium.

| Source type | What to look for | Where |
|---|---|---|
| Surveys / reviews | Existing maps of the field, taxonomy of approaches | arXiv, Google Scholar, Semantic Scholar |
| Primary papers | New methods, benchmark results | arXiv, conference proceedings |
| Benchmark leaderboards | Current best scores, what "SOTA" means concretely | Papers With Code, leaderboard sites |
| Engineering blogs | What works in production, real-world trade-offs | Company tech blogs, personal blogs |
| Documentation / specs | How mature tools actually behave | Official docs, RFCs, standards |
| Community discussion | Controversies, pitfalls, emerging work | HN, Reddit, forum threads |

Prefer recency for a fast-moving field, but rely on the seminal works that everyone cites.

## When to Stop Researching

Research converges when:
- New searches mostly surface sources already in the candidate list.
- The main approaches are stable and you can name them without checking notes.
- The top benchmark numbers and the dominant method are clear.

If after reasonable effort the field is too large to map fully, narrow the scope and say so explicitly in the brief rather than producing a shallow survey.

## Output Format

Write the research brief as markdown with this structure:

```markdown
---
topic: "<topic>"
audience: "<intended audience, or omit if unknown>"
status: draft
---

# Research Brief: <Topic>

## Scope

<One paragraph: the specific question investigated, what is in and out of scope, time bounds, and how "state of the art" is defined here.>

## Search Strategy

| Source type | Searched | Queries / notes |
|---|---|---|
| Surveys / reviews | Yes / No | <queries> |
| Primary papers | Yes / No | ... |
| Benchmark leaderboards | Yes / No | ... |
| Engineering blogs | Yes / No | ... |
| Documentation / specs | Yes / No | ... |
| Community discussion | Yes / No | ... |

## State of the Art

### Main Approaches

| Approach | Key idea | Representative work | Strengths | Limitations |
|---|---|---|---|---|
| <name> | <one line> | [Title](https://...) | <one line> | <one line> |

### Current Best Results

<Benchmarks, leaderboards, or numbers that define the state of the art today. Cite each.>

### Consensus and Disagreement

<Where the field agrees, and where it actively debates.>

## Key Sources

| Source | Type | Year | Authority signal | Credibility | Why it matters |
|---|---|---|---|---|---|
| [Title](https://...) | Paper / Blog / Docs / Standard | <year> | <the concrete signal: venue, org, citations, stars/forks, vendor status, sample size> | Strong / Moderate / Weak (+ qualifier) | <one line> |

Every source you keep must carry an authority signal and a credibility rating.
Rate vendor content and personal blogs as such rather than treating them as independent.
List any source cited in the body but not promoted here, and any claim with no link, with its rating so the gap is visible.

## Open Questions / Gaps

1. <What is unsolved, contested, or just emerging.>
2. ...

## Recommendation for the Article

**Core insight:** <The single argument the research best supports.>
**Angle:** <How to frame the article given the state of the art.>
**Sources to lead with:** <2-4 sources that should support the article.>
```

## Quality Criteria

**Strong research briefs:**
- Answer "what is the state of the art?" with specific names, numbers, and dates, not vague generalities
- Distinguish what is established from what is contested or emerging
- Trace claims to primary sources with verifiable links
- Rate every source's credibility from a concrete authority signal, and separate vendor or personal-blog claims from independent evidence
- Name the main approaches and how they compare
- Recommend a concrete angle the article can take

**Avoid:**
- Listing sources without explaining what each contributes
- Leaving a source's credibility unrated
- Presenting vendor, SEO, or personal-blog numbers as independent evidence
- Treating every source as equally important (triage strictly)
- Stating opinions or impressions as if they were sourced facts
- Stopping at the first page of search results without triangulating
- Em-dash sentence structures, use commas or parentheses instead
- The banned terms listed in `banned-terms.txt` at the skills library root (one per line), unless one is genuinely the most appropriate term

## Example Usage

**Scenario 1: Fast-moving research field**
Topic: "test-time compute in LLMs", scope: "last 18 months, accuracy on math reasoning".
Search arXiv and Papers With Code, find the dominant scaling approach and the top benchmark scores, note the open question of cost at inference.
Recommend an article framing test-time compute as the new scaling axis.

**Scenario 2: Established engineering topic**
Topic: "vector databases", scope: "what to choose in 2025, latency and recall trade-offs".
Search engineering blogs and docs, compare the main products and libraries on benchmarks, note the consensus on approximate nearest neighbor and the disagreement on indexing strategy.
Recommend a practical selection-guide article.

**Scenario 3: Conceptual explainer**
Topic: "CRDTs for collaborative editing", scope: "how they work and where they struggle".
Start from the seminal papers and a key survey, map the main algorithms, surface the gap around complexity and memory overhead.
Recommend an article that explains CRDTs and covers their limitations.

## Next Step

Run `/create-article`, passing the key sources from this brief and the recommended core insight and angle.
The research brief is the "relevant sources" prerequisite for `/create-article`, so the writing step can focus on structure, tone, and audience calibration rather than on finding material.

After the article is drafted, run `/review-article` to audit it for accuracy, sourcing, and clarity.

## Useful Commands Reference

| Command | Description |
|---|---|
| `WebSearch` | Search broadly across the web, blogs, and docs |
| `WebFetch` | Deep-read a candidate source and extract its claims |
| `/arxiv-article <url>` | Download, archive, and summarize a single arXiv paper |
| `grep` / codebase search | Find how the topic is handled in the local repository, if relevant |
| `gh search repos ...` | Search the organization's repositories for prior art or related work |